SpaceXAI shipped Grok 4.6 on Wednesday, a post-training refresh of Grok 4.5 aimed at long-running agents and agentic coding, priced at $2 per million input tokens and $6 per million output, roughly 60% below GPT-5.6 Sol at parity on the headline composite.
On the nine-benchmark Artificial Analysis Intelligence Index, Grok 4.6 scores 61, matching OpenAI’s GPT-5.6 Sol and trailing Anthropic’s Fable 5 Max at 62. Grok 4.5 scored 56. The lift came without a bigger base: SpaceXAI retained the 1.5 trillion-parameter V9 architecture and extended the supplemental training run, using Grok 4.5 itself to regenerate supervised fine-tuning trajectories across reasoning, agent harnesses, STEM, software engineering, and knowledge work, then filtering problematic traces with model-based checks. It’s a distillation-flavored strategy, and it’s now the cheapest way to move a leaderboard number.
The agentic picture is more uneven. Grok 4.6 posts 65.9% on DeepSWE v1.1, up from 54%, but still behind GPT-5.6 Sol Max at 73%. On Terminal-Bench v3.0 it lands at 26%, against roughly 34% for both Sol Max and Fable 5 Max. No independent evaluation has confirmed the figures. The model ships through Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare, with double included usage in Cursor and Grok Build for a week.
Elon Musk called it “objectively #1 when considering intelligence, speed & cost” on X. That framing, intelligence-per-dollar rather than intelligence outright, is the tell.
On SpaceX’s August 4 earnings call, Musk pointed to Grok 4.7 in three to four weeks at a target 2.1 trillion parameters, and Grok 5 before the end of 2026. The release cadence is now the product.